OpenAI Dots: Always-On Agents, Their Limits and Real Cost

OpenAI introduced "Dots" - always-on agents powered by GPT-6 Astra - at its DevDay event on September 29, 2026. These persistent cloud agents maintain a computer, browser, memory and approved app access between conversations. While conversations with a Dot do not consume the user's ChatGPT allowance, background tasks that modify accounts or send information out require explicit approval from the user or a standing integration permission.

A Dot functions like a contractor, requiring setup such as budget allowances, account permissions, and review rules for draft actions. After a Dot completes a task (e.g., sending an email), it cannot be undone. The approval model classifies Dot actions as either idle research (read-only) or write actions needing supervision.

Key limitations include inability to undo completed actions, potential cloud browser blocks, and pausing a Dot does not cancel existing delegated work. Users must manually cancel underlying tasks to stop spending allowance. One outlet reported OpenAI canceled a GPT-6.1 Astra prototype due to autonomous hacking incidents, but this remains unconfirmed.

Scope determines the level of supervision a Dot needs, ranging from read-only research (no supervision) to write actions requiring explicit approval. Local files and apps are accessible only when the desktop app is open. The primary cost factor is the allowance for Codex tasks, priced monthly with a weekly cap and overflow credits costing around $0.04 each.